Deep learning embedded latent class joint modelling of time-to-event and longitudinal data

Area of Science:

• Biostatistics
• Statistical modeling
• Machine learning in healthcare

Background:

• Joint modeling of longitudinal and time-to-event data is crucial for complex medical data analysis.
• Joint latent class models (JLCMs) offer prognostic insights by identifying latent subgroups.
• Traditional JLCMs use multinomial logistic regression for mixing proportions, potentially missing nonlinearities.

Purpose of the Study:

• To extend JLCMs by incorporating neural networks for modeling class-membership submodels.
• To introduce the neural-network JLCM (NN-JLCM) for enhanced flexibility and predictive power.
• To evaluate the performance of NN-JLCM through simulations and real-world data application.

Main Methods:

• Development of the neural-network JLCM (NN-JLCM) using an expectation-maximization-type algorithm.
• Application of neural networks to model mixing proportions in the class-membership submodel.
• Extensive simulations and real data analysis to assess model performance.

Main Results:

• NN-JLCM demonstrated improved predictive accuracy compared to traditional JLCMs.
• Enhanced class separation was observed with the NN-JLCM.
• The NN-JLCM maintained competitive overall model fit while offering greater flexibility.

Conclusions:

• The NN-JLCM provides a more flexible and accurate approach for joint modeling of longitudinal and time-to-event data.
• This novel method offers improved prognostic insights through better subgroup identification.
• The NN-JLCM represents a significant advancement in statistical modeling for complex biomedical research.